Through the Eyes of Emotion - Processed dataset

The Through the Eyes of Emotion processed dataset is a ready-to-train, curated version of the raw VR eye-tracking data, where all modalities are synchronized, segmented into fixed-length samples, normalized, and serialized into per-participant H5 files (processed_dataset/[processing_parameters]/Pxx.h5) with an accompanying index_info.json for fast indexing and loader integration.

Each H5 file contains a labels array (compound dtype with subject, label, and session_label) plus modality tensors: gaze_pupil feature matrices and eye0/eye1 frame sequences, all temporally aligned and standardized; samples are derived from the session timestamps in user_label.json, with consistent preprocessing (e.g., resizing, normalization, feature standardization) and removal of invalid/missing windows.

This release supports immediate use in supervised emotion recognition, multimodal fusion, and temporal sequence modeling, and is compatible with the provided training scripts and configurations for reproducible results; please cite “Through the Eyes of Emotion: A Multi‑faceted Eye Tracking Dataset for Emotion Recognition in Virtual Reality” when using this resource.